Alexander Embiricos and Roman Huitt of OpenAI argue that far from replacing engineers, AI is triggering a return to engineering's roots — "AI engineers are eating the world." They frame Codex as a layered, open stack (model, responses API, open-source harness, apps server, app-layer plugins) that OpenAI uses internally exactly as it ships to developers, and preview the GPT 5.6 series across cost-efficient (Terra, Luna), fast (Codex Spark, Cerebras at 750 tok/s), and frontier variants. The talk closes on "value maxing" — extracting real value from agents through cost, speed, and parallel cloud execution rather than "token maxing."
apps server, and open-source app-layer plugins — so "we're not building one system for OpenAI and a second simplified one for developers."AGENTS.md was named generically on purpose so other agents can adopt it; the harness defaults to OpenAI models but they aren't hardcoded, so open models can reuse the same agent loop.Roman and Alexander opened to a room of over 7,000 AI engineers, invoking the World's Fair as a place where impossible ideas became visible "by building it in public." They pushed back directly on the narrative that engineers are going away as coding gets abstracted: "in fact, we think it's quite the opposite."
Their framing: software ate the world, then AI ate software, and now the AI engineers are eating the world. Engineering, they argued, was never about writing code — it's about solving problems by combining the latest science with design, taste, judgment, and imagination. In that sense it's "not the end of engineering" but "a return to the roots of engineering."
Alexander traced the successive "mind-blowing" phases: completion, inline prediction, Command-K edits that couldn't test their work, models that began testing their own work, and now models "taking on long hard goals until they're done." He recalled using o1 preview at Dev Day 2024 to build a drone interface — where the model couldn't run or verify its own code, so he had to cross his fingers on the live demo — versus 2025, when self-testing models could control an entire camera and lighting system live.
The pace of product progress mirrors the model progress. Favorite recent ships include the Codex app, goal mode, and remote. The key shift: agents can now do "any task that you can do on your own computer," helping with what happens before and after the coding — connecting an agent to why work must be done helps it begin work, and connecting it to review and deploy helps it land work.
Looking to OpenAI's mission of "AGI that benefits all of humanity," Alexander said the goal is "squarely not to automate engineers" but to build a product shape that "maximally empowers engineers." Drawing on sci-fi, he described two modalities: chat ("I think chat is underrated") as a single entity you can ask for help with anything anywhere, and a powerful collaborative UI for when you want to inspect, steer, or shape things yourself.
The analogy is working with a team: most of the time you talk and "let them cook," then occasionally dig all the way into the weeds together. This drove the Codex app — a simple chat interface for coding or anything else, where you can point at a specific thing and either ask the model to change it or change it yourself. Alexander noted the app was "quite a controversial project to start," with skeptics who swore they'd never leave the terminal, Vim, or Emacs now using it. His take: you can't build a collaborative interface in a CLI, and in an IDE "the order is wrong" — you should chat first and dig into code when you need to.
Roman stressed that "Codex cannot be a closed product that only OpenAI can improve," so it's designed as layers anyone can build on. It starts with the model, accessed through the responses API — the exact same models and API OpenAI uses to build the Codex app. New Codex needs get baked into the API first (e.g., compaction for long-running task contexts) so developer agents get the same primitives.
The Codex harness is open source — inspectable, forkable, adaptable — and AGENTS.md was named so other agents could adopt it. Models are the default but not hardcoded, so open models can reuse the same agent loop, and the harness is used in post-training so models learn to call tools in an open-source environment. Above that sits the open-source apps server (the real path OpenAI uses for its VS Code extension and Codex app, not a community adapter) and extensible app-layer primitives like the in-app browser and plugins — including open-source role-specific plugins for data science and design. The Open Code team reused the reference implementation; Tuma (aka Dimillian) built Codex Monitor via the apps server before the app launched and now builds Codex for iOS on the team.
Alexander reframed the goal from "token maxing" to value maxing — getting real value out of agents. On cost efficiency: frontier GPT 5.6 tops terminal bench, while GPT 5.6 Terra brings GPT 5.5-level intelligence at half the cost, and Luna beats notable models at just $1 per million input and $6 per million output tokens.
On speed, after GPT 5.3 Codex Spark, GPT 5.6 running on Cerebras delivers frontier intelligence at 750 tokens/second — "a pretty substantial PR in like 10 seconds" — enabling an agent to try five or six approaches in parallel and pick the best. The vision extends to running many isolated tasks so you can close your laptop: Codex Cloud (their first major launch, with upgrades coming) should eventually erase the awkward local-vs-cloud distinction, leaving just an agent that figures out which environment fits the work. Citing a Theo tweet, Alexander predicted this arrives "much sooner than six months," before teasing a special guest instead of the usual live demo.
"Software ate the world and then AI ate software. But now what we're here to say is that the AI engineers are eating the world."
"Engineering was never about writing code. Engineering has always been about solving problems for yourself and for other people as well."
"We used to ship a new model every 15 months or so, and now it's about roughly every six weeks."
"We're not building one system for OpenAI and a second system that's simplified for developers. At every layer, we actually use the thing that we give to you."
"This is kind of like having a pretty substantial PR written in like 10 seconds."
| Time | Topic |
|---|---|
| 00:30 | Codex team takes the stage |
| 02:02 | AI engineers are eating the world |
| 03:33 | From completion to long-horizon goals |
| 07:40 | Chat plus a collaborative UI |
| 10:42 | Codex as an open, layered stack |
| 15:17 | Value maxing: cost, speed, cloud |
AGENTS.md, open-source Codex harness, responses API, and open-source apps server